Fully convolutional neural network (FCN) for semantic segmentation with tensorflow.
This is a simple implementation of a fully convolutional neural network (FCN). The net is based on fully convolutional neural net described in the paper Fully Convolutional Networks for Semantic Segmentation. The code is based on FCN implementation by Sarath Shekkizhar with MIT license but replaces the VGG19 encoder with VGG16 encoder. The net is initialized using the pre-trained VGG16 model by Marvin Teichmann. An improved version of this net in pytorch is given here
This is an old model for a New more accurate version of the net focused on recognition of materials in transparent vessels see this link
Details input/output
The input for the net is RGB image (Figure 1 right). The net produces pixel-wise annotation as a matrix in the size of the image with the value of each pixel corresponding to its class (Figure 1 left).
Figure 1) Semantic segmentation of image of liquid in glass vessel with FCN. Red=Glass, Blue=Liquid, White=Background
Requirements
This network was run with Python 3.6 Anaconda package and Tensorflow 1.1. The training was done using Nvidia GTX 1080, on Linux Ubuntu 16.04.
Setup
1) Download the code from the repository. 2) Download a pre-trained vgg16 net and put in the /Model_Zoo subfolder in the main code folder. A pre-trained vgg16 net can be download from here[https://drive.google.com/file/d/0B6njwynsu2hXZWcwX0FKTGJKRWs/view?usp=sharing] or from here [ftp://mi.eng.cam.ac.uk/pub/mttt2/models/vgg16.npy]
Tutorial
Instructions for training (in TRAIN.py):
In: TRAIN.py 1) Set folder of the training images in Train_Image_Dir 2) Set folder for the ground truth labels in Train_Label_DIR 3) The Label Maps should be saved as png image with the same name as the corresponding image and png ending 4) Download a pretrained vgg16 model and put in model_path (should be done automatically if you have internet connection) 5) Set number of classes/labels in NUM_CLASSES 6) If you are interested in using validation set during training, set UseValidationSet=True and the validation image folder to Valid_Image_Dir and set the folder with ground truth labels for the validation set in Valid_Label_Dir
Instructions for predicting pixelwise annotation using trained net (in Inference.py)
In: Inference.py 1) Make sure you have trained model in logs_dir (See Train.py for creating trained model) 2) Set the Image_Dir to the folder where the input images for prediction are located. 3) Set the number of classes in NUM_CLASSES 4) Set folder where you want the output annotated images to be saved to Pred_Dir 5) Run script
Evaluating net performance using intersection over union (IOU):
In: Evaluate_Net_IOU.py 1) Make sure you have trained model in logs_dir (See Train.py for creating trained model) 2) Set the Image_Dir to the folder where the input images for prediction are located 3) Set folder for ground truth labels in Label_DIR. The Label Maps should be saved as png image with the same name as the corresponding image and png ending 4) Set number of classes number in NUM_CLASSES 5) Run script
Supporting data sets
The net was tested on a dataset of annotated images of materials in glass vessels. This dataset can be downloaded from here
MIT Scene Parsing Benchmark with over 20k pixel-wise annotated images can also be used for training and can be download from here
Trained model
Glass and transparent vessel recognition trained model
Liquid Solid chemical phases recognition in transparent glassware trained model